Spatial morphoproteomic features predict disease states from tissue architectures

Understanding how immune cells organize within tissue microenvironments is essential for interpreting disease responses in spatial proteomics data. We introduce SNOWFLAKE, a graph neural network pipeline that integrates single-cell protein expression and morphological features to predict disease status from lymphoid follicles. Using a pediatric COVID-19 dataset, SNOWFLAKE outperformed conventional machine learning and deep learning approaches in classifying infection status. By incorporating morphology into graph edge features, SNOWFLAKE enables the identification of spatially organized subgraphs associated with disease. These subgraphs, derived from single-cell neighborhoods, display clear distinctions between COVID-positive and negative cases and reveal interpretable cellular motifs. SNOWFLAKE's ability to extract meaningful subgraph embeddings highlights its value in understanding immune architecture and its alterations in disease. The approach generalizes across tissue types, including breast cancer and tertiary lymphoid structures, underscoring its utility for spatial systems biology and biomarker discovery from multiplex imaging data.